Paper Abstract and Keywords |
Presentation |
2017-06-25 11:25
Cost-sensitive Bayesian optimization for multiple objectives and its application to material science Tomohiro Yonezu (NITech), Tomoyuki Tamura, Ryo Kobayashi (NITech/NIMS), Ichiro Takeuchi (NITech/NIMS/RIKEN), Masayuki Karasuyama (NITech/NIMS/JST) IBISML2017-10 |
Abstract |
(in Japanese) |
(See Japanese page) |
(in English) |
We consider solving a set of black-box optimization problems in which each problem has a similar objective function each other.
For example, in the crystal structure search problem in material science, identifying minimum energy points in a similar multiple energy surfaces generated by different types of crystals is an important problem.
Bayesian optimization is a standard approach to the black-box optimization by which single objective function can be efficiently explored.
In this study, we extend Gaussian process in Bayesian optimization to multi-task Gaussian process to deal with multiple objective functions efficiently.
By introducing between-task similarity by a task kernel function, the optimization process can be faster than applying single task Bayesian optimization separately.
Furthermore, we discuss cost-sensitive scenario for multiple objective functions.
The entire exploration cost can be decreased by constructing an accurate Gaussian process model using lower cost samples before searching higher cost samples because of their similarity.
In our experiments, we verify effectiveness of our approach based on synthetic problems and an application to an energy search problem of crystal structures, called grain-boundary.
We will show that, in the grain-boundary search, there exist multiple objective functions with largely different sample cost. |
Keyword |
(in Japanese) |
(See Japanese page) |
(in English) |
Gaussian Process / Multi-task Machine learning / Materials Informatics / / / / / |
Reference Info. |
IEICE Tech. Rep., vol. 117, no. 110, IBISML2017-10, pp. 207-213, June 2017. |
Paper # |
IBISML2017-10 |
Date of Issue |
2017-06-17 (IBISML) |
ISSN |
Print edition: ISSN 0913-5685 Online edition: ISSN 2432-6380 |
Copyright and reproduction |
All rights are reserved and no part of this publication may be reproduced or transmitted in any form or by any means, electronic or mechanical, including photocopy, recording, or any information storage and retrieval system, without permission in writing from the publisher. Notwithstanding, instructors are permitted to photocopy isolated articles for noncommercial classroom use without fee. (License No.: 10GA0019/12GB0052/13GB0056/17GB0034/18GB0034) |
Download PDF |
IBISML2017-10 |
|